An Evaluation of Pre-Trained Deep Learning Algorithms for Diabetic Retinopathy Disease Identification Using Retinal Fundus Images

Abstract

A persistently elevated blood sugar level can lead to diabetic retinopathy (DR), an eye disorder that damages retinal tissue by obstructing and bleeding retinal capillaries. Usually, it causes blindness. The risk and severity of DR can be reduced by early identification. Diabetic retinopathy is difficult to detect and predict with reliability and accuracy. In this paper, a completely accurate deep learning model for Diabetic Retinopathy identification is developed. Utilizing a transfer learning (TL) strategy, pre-trained models with a pooling layer, dense layer, and suitable dropout layer at the bottom were used, including ResNet50, InceptionV3, Alexnet, and VGG19.  To minimize overfitting, data augmentation along with regularization were carried out. The DL systems had been trained and evaluated on the Asia Pacific Tele-Ophthalmology Society (APTOS) datasets. The robustness of the chosen models was demonstrated by the calculation of testing accuracy and performance metrics, including precision, recall, and F1 score. The AlexNet model shows its greatest testing accuracy at 98.66%. Additionally, the obtained evolution metrics reinforced our obtained results. Additionally, AlexNet has a limited number of the layers, which reduces training time as well as computational complexity.